SYNTHETIC-1 Release: Two Million Collaboratively Generated Reasoning Traces from Deepseek-R1
Blog post from Prime Intellect
SYNTHETIC-1, released from Deepseek-R1, is the largest open reasoning dataset collaboratively generated worldwide, featuring reasoning traces for tasks in math, coding, and science, verified for accuracy by task-specific validators. This dataset includes both correct and incorrect reasoning traces enriched with metadata and offers a curated Supervised Fine-Tuning (SFT) subset with 900k samples, marking it the largest from R1, alongside a preference tuning dataset derived from varying response correctness. The dataset's effectiveness in teaching models reasoning through supervised fine-tuning is demonstrated by the SYNTHETIC-1-SFT-7B model, which significantly enhances reasoning performance. The dataset was built using Genesys, an open-source library facilitating the development and integration of verifiers for synthetic data generation and reinforcement learning, supporting diverse tasks with unique verification methods. The total dataset encompasses 2 million responses, with a post-processed SFT dataset and preference dataset available for further development. Future plans involve a globally distributed collaborative reinforcement learning run, aiming to train a state-of-the-art reasoning model with significant parameter expansion.
No tracked trend matches for this post yet.
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.